CuArray only supports element types that are stored inline?

Specific Domains GPU
, , ,
1

Hello,

I would like to define, a mutable structure with a field of type ::CuArray{an_other_mutable_structure} but I get this error:

CuArray only supports element types that are stored inline

I don’t understand what it means, but it comes from the constructor of CuArray:

function CuArray{T,N}(::UndefInitializer, dims::Dims{N}) where {T,N}
    Base.allocatedinline(T) || error("CuArray only supports element types that are stored inline")
....

Here is my code:

Tf = Float64
Ti = Int64
mutable struct node_struct
    x::Tf
    y::Tf
    xr::Tf
    zr::Tf
    vx::Tf
    vz::Tf
    Fx::Tf
    Fz::Tf  
    node_struct() = new(0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0)
end

mutable struct IBM_object
    is_wall::Bool
    num_nodes::Ti
    radius::Tf
    stiffness::Tf
    nodes::CuArray{node_struct}
    IBM_object(number) = new(true, number, 0, 1, CuArray{node_struct}(undef, number))
end

At the beginning I was thinking that it could be solved with Adapt, but with Adapt.@adapt_structure node_struct before the definition of the second structure nothing changed.

Best,

Ludo

2

You’re not allowed to store mutable structs as the element types of CuArrays (or any GPUArray) because they end up getting stored as pointers, and being heap-allocated and managed. This is incompatible with GPU execution, because we don’t have access to Julia’s runtime+GC while running on the GPU.

Instead, I’d use an immutable struct, and use something like Setfield.jl to read-modify-write array values.